China Didn’t Wait For Better AI Chips

Jensen Huang’s interview with Dwarkesh Patel this week is worth your time, especially the extended exchange from 57:00 on whether the US should selling AI chips to China.

Jensen’s take on this question is straightforward. Rather than slowing China’s AI development, restricting Nvidia sales has accelerated the development of a parallel, non-American stack.

Jensen explains why that’s been bad for Nvidia (obviously), while also being bad for the US. I won’t get into the geopolitics of it, but sitting in Beijing, I can attest that Chinese AI development has not been standing still while waiting for better hardware. The constraints have produced a different kind of innovation, of which DeepSeek is the most visible example. Chinese researchers had to optimize aggressively for efficiency, and Chinese models now deliver high performance at a fraction of the compute cost. And as Jensen points out, compute in China is already cheaper and more abundant than the export control narrative suggests.

So with that efficiency, over here we’ve got tons of models that are already deeply embedded in applications in ways that Western AI largely isn’t. Alibaba, JD, Pinduoduo, and Douyin have integrated large models directly into their e-commerce stacks, covering product discovery, customer service, live-stream commerce, logistics optimization, and more. Enterprise adoption is similarly advanced. The gap between model capability and real-world deployment that still exists in the West has largely been closed here.

This connects to the broader point I’ve been making from the consumer and commercial side for years. China has one of the most sophisticated technology ecosystems in the world. The platforms, the supply chains, the pace of AI adoption and application at scale. Companies like Alibaba cover every aspect of the AI stack.

Highly recommend listening to Jensen’s full thoughts.